Re: [PATCH v7 2/4] dmaengine: Forward slave device pointer to of_xlate callback

From: Marek Szyprowski
Date: Thu Feb 09 2017 - 04:55:14 EST


Hi Vinod,

On 2017-02-09 05:11, Vinod Koul wrote:
On Thu, Jan 26, 2017 at 03:43:05PM +0100, Marek Szyprowski wrote:
On 2017-01-25 14:12, Lars-Peter Clausen wrote:
On 01/25/2017 11:28 AM, Marek Szyprowski wrote:
Add pointer to slave device to of_dma_xlate to let DMA engine driver
to know which slave device is using given DMA channel. This will be
later used to implement non-irq-safe runtime PM for DMA engine driver.
of_dma_xlate() is used to translate from a OF phandle and a specifier to a
DMA channel. On one hand this does not necessarily mean that the channel is
actually going to be used by the slave that called the xlate function.
Modifying the driver state when a lookup of the channel is done is a
layering violation. And this approach is also missing a way to disassociate
a slave from a DMA channel, e.g. when it is no longer used.

On the other hand there are other mechanisms to translate between some kind
of firmware handle to a DMA channel which are completely ignored here.

So this approach does not work. This is something that needs to be done at
the dmaengine level, not a the firmware resource translation level. And it
needs a matching method that is called when the channel is disassociated
>from a device, when the device no longer uses the DMA channel.

Frankly I agree that of_dma_xlate() should only return the requested channel
to the dmaengine core and do not do any modification in the the
driver state.
True..

However the current dma engine design and implementation breaks this rule.
Please check the drivers - how do they implement of_xlate callback. They
usually call dma_get_any_slave_channel, dma_get_slave_channel or
__dma_request_channel there, which in turn calls dma_chan_get, which then
calls back to device_alloc_chan_resources callback. Some of the drivers also
do a hardware configuration or other resource allocation in of_xlate.
This is a bit messy design and leave no place in the core to set
slave device
before device_alloc_chan_resources callback, where one would expect to have
it already set.
We shouldn't be doing much at this stage. We operate on a channel, so the
channel is returned to the client. We need to do these HW configurations
when the channel has to be prepred for a txn.

IMHO, any HW configurations should be done in alloc_chan_resources callback and
it would be best if a pointer of slave device will come as a parameter to it.

The best place to add new calls to the dmaengine drivers to set slave device
would be just before device_alloc_chan_resources(), what in turn means that
the current dmaengine core should do in dma_chan_get(). This would
require to
forward the slave device pointer via even more layers including the of_xlate
callback too. IMHO this is not worth the effort.

DMA engine core and API definitely needs some cleanup. During such cleanup
the slave device pointer might be moved out of xlate into separate callback
when the core gets ready for such operation.
Yes agreed on that, plus the runtime handling needs to be built in, right
now the APIs dont work well with it, we disucssed these during the KS and
this goes without saying, patches are welcome :)

Okay, so what is the conclusion? Do you want me to do the whole rework of dma
engine core to get this runtime pm patchset for pl330 merged??? Is there any
roadmap for this rework prepared, so I can at least take a look at the amount
of work needed to be done?

I'm rather new to dma engine framework and I only wanted to fix pl330 driver
not to block turning off the power domain on Exynos5422/5433 SoCs.

I can also check again if there is any other way to find the slave device in
alloc_chan_resources, like for example scanning the device tree for phandles,
to avoid changing dmaengine core as this turned out to be too problematic
before one will do the proper dma engine core rework.

> ...

Best regards
--
Marek Szyprowski, PhD
Samsung R&D Institute Poland